Andreas,
I don't know anything about sqm gradients. All I can say is that my
TNCG minimizer behaves in ways that are indicative of gradients to be
off at least to the extent that they cannot be used reliably in finite
difference Hessian * vector products where H(x) * v = (g(x) -
g(x+eps*v)) / eps.
Is there a way to produce accurate sqm gradients even if it is
expensive, for debugging purposes?
Istvan
Quoting Andreas Goetz <agoetz.sdsc.edu>:
> Hi,
>
> How inaccurate are the gradients?
> How did you test this? (I would test against finite difference gradients
> but these are obviously limited in accuracy)
>
> All the best,
> Andy
>
> istvan.kolossvary.hu wrote:
>> Dave,
>>
>> I looked at this as much as I could. There is nothing wrong with XMIN
>> settings. The problem seems to be that as you indicated sqm gradients
>> are not accurate. It bothers lbfgs less, but it kills tncg. The reason
>> being that tncg relies on numerous Hess x some_vector products and these
>> are computed via finite difference using gradients. Unfortunately, if
>> gradients are off (it is probably very difficult to quantify how much
>> off is significant) the product won't be just inaccurate, it would just
>> not be the product of the Hessian with something. Therefore, the
>> conjugate gradient part of tncg runs amok and will never converge. I
>> guess, at this point we are stuck with lbfgs for sqm. I tried larger
>> lbfgs memory, but it didn't change things too much. What is the reason
>> for the inaccurate gradients? Can it be fixed. Even if it is way too
>> slow, just to check that minimization works better? I'd be glad to look
>> into this some more, but right now I don't have any suggestion.
>>
>> Istvan
>>
>> Quoting case <case.biomaps.rutgers.edu>:
>>
>>> Hi Istvan:
>>>
>>> About minimization methods in sqm, look at the test case in
>>> amber11/test/antechamber/sustiva. (I've attached the sqm input; you
>>> can just type 'sqm -i sqm.in -o output-file').
>>>
>>> With LBFGS, this input runs for 3650 cycles, reaching a final energy
>>> of -120.503799 and a gradient of 0.000498. The energy gets to within 0.1
>>> kcal/mol of its final value after 2270 cycles.
>>>
>>> With TCNG, the minimization exits after 10,000 cycles, with a energy of
>>> -120.3472 and a gradient of 0.001713. It gets to within 0.1 kcal/mol
>>> of its final energy after 860 cycles, but this is still 0.17 kcal/mol
>>> higher
>>> than the state found by LBFGS.
>>>
>>> So: this might be the result of gradients that are slightly wrong, or
>>> it might
>>> indicate that one of the parameters defining the TCNG scheme is not set
>>> correctly in sqm. (Or, something else....)
>>>
>>> ...regards...dave
>>>
>>>
>>> _______________________________________________
>>> AMBER-Developers mailing list
>>> AMBER-Developers.ambermd.org
>>> http://lists.ambermd.org/mailman/listinfo/amber-developers
>>>
>>
>>
>>
>>
>> _______________________________________________
>> AMBER-Developers mailing list
>> AMBER-Developers.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Tue Dec 15 2009 - 08:00:02 PST
